Teaching Machines to Understand Baseball Games: Large Scale Baseball Video Database for Multiple Video Understanding Tasks

Young Hwi Kim*

Kyungmin Kim*

Department of Computer Science

Yonsei University

* indicates equal contribution


A major obstacle in teaching machines to understand videos is the lack of training data, as creating temporal annotations for long videos requires a huge amount of human effort. To this end, we introduce a new large-scale baseball video dataset called the BBDB, which is produced semi-automatically by using play-by-play texts available online. The BBDB contains 4200 hours of baseball game videos with 400k temporally annotated activity segments. The new dataset has several major challenging factors compared to other datasets: 1) the dataset contains a large number of visually similar segments with different labels. 2) It can be used for many video understanding tasks including video recognition, localization, text-video alignment, video highlight generation, and data imbalance problem. To observe the potential of the BBDB, we conducted extensive experiments by running many different types of video understanding algorithms on our new dataset.

Paper: [PDF]


License for any software and data in this page is available here.


If you cannot see files below, click here. Click here to see the password to unzip the file.

Please prepare 4TB of free space before starting download.

BBDB v0.9 contains information used in experiments section of the paper. BBDB v1.0 will be similar in size, but new games will be added with more thorough validation. Along with BBDB, corresponding baseball highlight videos and their annotations will be provided when ready.


This json file is structured as below, as an example:

  "version": 0.9,
  "database": {
    "20170802AABB02017": { // Unique ID for each game
      "duration": 12068.04, // duration in seconds
      "subset": "validation",
      "resolution": "852x480",
      "fps": 29.97,
      "pageUrl": "http://link.to.page",
      "videoUrl": "http://link.to.video",
      "annotations": [
          "segment": [
            782.039, // start point of a segment, in second
            787.039  // end point of a segment
          "labelIndex": 0,
          "label": "Ball"
          "segment": [
          "labelIndex": 1,
          "label": "Strike"
          "segment": [
          "labelIndex": 0,
          "label": "Ball"
        }, ...
    }, ...


This is a helper python script to download BBDB as fast as possible. With the highest speed (~200KB/s per connection), download may take 7-10 days. The script supports a simple method to continue download after pausing, but be careful doing so since the method may not be perfect.

You need aria2 (version ≥ 1.33.0) to be installed, or a binary executable file is required. When using a binary, you can specify the path with -b option.

usage: bbdb_downloader.py [-h] [-v TARGET_DIR] [-i JSON_FILE] [-b ARIA2_BINARY]

optional arguments:
  -h, --help            show this help message and exit
  -v TARGET_DIR, --target-dir TARGET_DIR, 
     Target directory to download BBDB dataset (default: ./)
  -i JSON_FILE, --input-json-file JSON_FILE
     BBDB JSON file (default: ./bbdb.v0.9.min.json)
  -b ARIA2_BINARY, --aria2-binary-path ARIA2_BINARY
     Specify path to aria2 binary (default: aria2c)


M. Shim*, Y. H. Kim*, K. Kim*, S. J. Kim, "Teaching Machines to Understand Baseball Games: Large Scale Baseball Video Database for Multiple Video Understanding Tasks", Proc. European Conference on Computer Vision (ECCV), 2018 (* indicates equal contribution)

  title={{T}eaching Machines to Understand Baseball Games: {L}arge Scale Baseball Video Database for Multiple Video Understanding Tasks},
  author={Shim, Minho and Kim, Young Hwi and Kim, Kyungmin and Kim, Seon Joo},
  booktitle={European Conference on Computer Vision (ECCV)},